Skip to content

model: correctly support input vision for deepseek4 - #28154

Merged
ngxson merged 2 commits into
masterfrom
xsn/dsv4_vision_prob_bias
Sep 2, 2026
Merged

model: correctly support input vision for deepseek4#28154
ngxson merged 2 commits into
masterfrom
xsn/dsv4_vision_prob_bias

Conversation

@ngxson

@ngxson ngxson commented Sep 1, 2026

Copy link
Copy Markdown
Collaborator

Overview

Important

Text model need to be reconverted

Stack on top of #28133

The DeepSeek-V4-Flash-Vision-Exp requires two extra things:

  • specific exp_probs bias for vision input
  • do not apply SWA when non_causal is set

Requirements

@github-actions github-actions Bot added model Model specific mtmd Related to multimodal functionality (video/image/audio) conversion labels Sep 1, 2026
@ngxson
ngxson marked this pull request as ready for review September 1, 2026 09:59
@ngxson
ngxson requested review from a team, CISC and ggerganov as code owners September 1, 2026 09:59
@nikwen

nikwen commented Sep 1, 2026

Copy link
Copy Markdown
Member

Looks like this has a merge conflict.

@ngxson
ngxson force-pushed the xsn/dsv4_vision_prob_bias branch from 5786559 to 316ed8a Compare September 2, 2026 14:46
@ngxson
ngxson requested review from a team, JohannesGaessler and danbev as code owners September 2, 2026 14:46
@ngxson
ngxson changed the base branch from xsn/dsv4_vision to master September 2, 2026 14:47
@ngxson
ngxson removed request for a team September 2, 2026 14:47
@ngxson
ngxson removed request for a team, JohannesGaessler and danbev September 2, 2026 14:47
@ngxson
ngxson merged commit 9400c89 into master Sep 2, 2026
3 checks passed
Lawlietr pushed a commit to Lawlietr/llama.cpp that referenced this pull request Sep 3, 2026
* model: correctly support input vision for deepseek4

* nits
Te-eMster pushed a commit to Te-eMster/mx-llama.cpp that referenced this pull request Sep 4, 2026
* model: correctly support input vision for deepseek4

* nits
fewtarius pushed a commit to fewtarius/CachyLLama that referenced this pull request Sep 5, 2026
* model: correctly support input vision for deepseek4

* nits
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

conversion model Model specific mtmd Related to multimodal functionality (video/image/audio)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants